async-thread.c 9.1 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364
  1. /*
  2. * Copyright (C) 2007 Oracle. All rights reserved.
  3. * Copyright (C) 2014 Fujitsu. All rights reserved.
  4. *
  5. * This program is free software; you can redistribute it and/or
  6. * modify it under the terms of the GNU General Public
  7. * License v2 as published by the Free Software Foundation.
  8. *
  9. * This program is distributed in the hope that it will be useful,
  10. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  12. * General Public License for more details.
  13. *
  14. * You should have received a copy of the GNU General Public
  15. * License along with this program; if not, write to the
  16. * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  17. * Boston, MA 021110-1307, USA.
  18. */
  19. #include <linux/kthread.h>
  20. #include <linux/slab.h>
  21. #include <linux/list.h>
  22. #include <linux/spinlock.h>
  23. #include <linux/freezer.h>
  24. #include "async-thread.h"
  25. #include "ctree.h"
  26. #define WORK_DONE_BIT 0
  27. #define WORK_ORDER_DONE_BIT 1
  28. #define WORK_HIGH_PRIO_BIT 2
  29. #define NO_THRESHOLD (-1)
  30. #define DFT_THRESHOLD (32)
  31. struct __btrfs_workqueue {
  32. struct workqueue_struct *normal_wq;
  33. /* List head pointing to ordered work list */
  34. struct list_head ordered_list;
  35. /* Spinlock for ordered_list */
  36. spinlock_t list_lock;
  37. /* Thresholding related variants */
  38. atomic_t pending;
  39. int max_active;
  40. int current_max;
  41. int thresh;
  42. unsigned int count;
  43. spinlock_t thres_lock;
  44. };
  45. struct btrfs_workqueue {
  46. struct __btrfs_workqueue *normal;
  47. struct __btrfs_workqueue *high;
  48. };
  49. static void normal_work_helper(struct btrfs_work *work);
  50. #define BTRFS_WORK_HELPER(name) \
  51. void btrfs_##name(struct work_struct *arg) \
  52. { \
  53. struct btrfs_work *work = container_of(arg, struct btrfs_work, \
  54. normal_work); \
  55. normal_work_helper(work); \
  56. }
  57. BTRFS_WORK_HELPER(worker_helper);
  58. BTRFS_WORK_HELPER(delalloc_helper);
  59. BTRFS_WORK_HELPER(flush_delalloc_helper);
  60. BTRFS_WORK_HELPER(cache_helper);
  61. BTRFS_WORK_HELPER(submit_helper);
  62. BTRFS_WORK_HELPER(fixup_helper);
  63. BTRFS_WORK_HELPER(endio_helper);
  64. BTRFS_WORK_HELPER(endio_meta_helper);
  65. BTRFS_WORK_HELPER(endio_meta_write_helper);
  66. BTRFS_WORK_HELPER(endio_raid56_helper);
  67. BTRFS_WORK_HELPER(rmw_helper);
  68. BTRFS_WORK_HELPER(endio_write_helper);
  69. BTRFS_WORK_HELPER(freespace_write_helper);
  70. BTRFS_WORK_HELPER(delayed_meta_helper);
  71. BTRFS_WORK_HELPER(readahead_helper);
  72. BTRFS_WORK_HELPER(qgroup_rescan_helper);
  73. BTRFS_WORK_HELPER(extent_refs_helper);
  74. BTRFS_WORK_HELPER(scrub_helper);
  75. BTRFS_WORK_HELPER(scrubwrc_helper);
  76. BTRFS_WORK_HELPER(scrubnc_helper);
  77. static struct __btrfs_workqueue *
  78. __btrfs_alloc_workqueue(const char *name, int flags, int max_active,
  79. int thresh)
  80. {
  81. struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
  82. if (unlikely(!ret))
  83. return NULL;
  84. ret->max_active = max_active;
  85. atomic_set(&ret->pending, 0);
  86. if (thresh == 0)
  87. thresh = DFT_THRESHOLD;
  88. /* For low threshold, disabling threshold is a better choice */
  89. if (thresh < DFT_THRESHOLD) {
  90. ret->current_max = max_active;
  91. ret->thresh = NO_THRESHOLD;
  92. } else {
  93. ret->current_max = 1;
  94. ret->thresh = thresh;
  95. }
  96. if (flags & WQ_HIGHPRI)
  97. ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
  98. ret->max_active,
  99. "btrfs", name);
  100. else
  101. ret->normal_wq = alloc_workqueue("%s-%s", flags,
  102. ret->max_active, "btrfs",
  103. name);
  104. if (unlikely(!ret->normal_wq)) {
  105. kfree(ret);
  106. return NULL;
  107. }
  108. INIT_LIST_HEAD(&ret->ordered_list);
  109. spin_lock_init(&ret->list_lock);
  110. spin_lock_init(&ret->thres_lock);
  111. trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
  112. return ret;
  113. }
  114. static inline void
  115. __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
  116. struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name,
  117. int flags,
  118. int max_active,
  119. int thresh)
  120. {
  121. struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
  122. if (unlikely(!ret))
  123. return NULL;
  124. ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI,
  125. max_active, thresh);
  126. if (unlikely(!ret->normal)) {
  127. kfree(ret);
  128. return NULL;
  129. }
  130. if (flags & WQ_HIGHPRI) {
  131. ret->high = __btrfs_alloc_workqueue(name, flags, max_active,
  132. thresh);
  133. if (unlikely(!ret->high)) {
  134. __btrfs_destroy_workqueue(ret->normal);
  135. kfree(ret);
  136. return NULL;
  137. }
  138. }
  139. return ret;
  140. }
  141. /*
  142. * Hook for threshold which will be called in btrfs_queue_work.
  143. * This hook WILL be called in IRQ handler context,
  144. * so workqueue_set_max_active MUST NOT be called in this hook
  145. */
  146. static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
  147. {
  148. if (wq->thresh == NO_THRESHOLD)
  149. return;
  150. atomic_inc(&wq->pending);
  151. }
  152. /*
  153. * Hook for threshold which will be called before executing the work,
  154. * This hook is called in kthread content.
  155. * So workqueue_set_max_active is called here.
  156. */
  157. static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
  158. {
  159. int new_max_active;
  160. long pending;
  161. int need_change = 0;
  162. if (wq->thresh == NO_THRESHOLD)
  163. return;
  164. atomic_dec(&wq->pending);
  165. spin_lock(&wq->thres_lock);
  166. /*
  167. * Use wq->count to limit the calling frequency of
  168. * workqueue_set_max_active.
  169. */
  170. wq->count++;
  171. wq->count %= (wq->thresh / 4);
  172. if (!wq->count)
  173. goto out;
  174. new_max_active = wq->current_max;
  175. /*
  176. * pending may be changed later, but it's OK since we really
  177. * don't need it so accurate to calculate new_max_active.
  178. */
  179. pending = atomic_read(&wq->pending);
  180. if (pending > wq->thresh)
  181. new_max_active++;
  182. if (pending < wq->thresh / 2)
  183. new_max_active--;
  184. new_max_active = clamp_val(new_max_active, 1, wq->max_active);
  185. if (new_max_active != wq->current_max) {
  186. need_change = 1;
  187. wq->current_max = new_max_active;
  188. }
  189. out:
  190. spin_unlock(&wq->thres_lock);
  191. if (need_change) {
  192. workqueue_set_max_active(wq->normal_wq, wq->current_max);
  193. }
  194. }
  195. static void run_ordered_work(struct __btrfs_workqueue *wq)
  196. {
  197. struct list_head *list = &wq->ordered_list;
  198. struct btrfs_work *work;
  199. spinlock_t *lock = &wq->list_lock;
  200. unsigned long flags;
  201. while (1) {
  202. spin_lock_irqsave(lock, flags);
  203. if (list_empty(list))
  204. break;
  205. work = list_entry(list->next, struct btrfs_work,
  206. ordered_list);
  207. if (!test_bit(WORK_DONE_BIT, &work->flags))
  208. break;
  209. /*
  210. * we are going to call the ordered done function, but
  211. * we leave the work item on the list as a barrier so
  212. * that later work items that are done don't have their
  213. * functions called before this one returns
  214. */
  215. if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
  216. break;
  217. trace_btrfs_ordered_sched(work);
  218. spin_unlock_irqrestore(lock, flags);
  219. work->ordered_func(work);
  220. /* now take the lock again and drop our item from the list */
  221. spin_lock_irqsave(lock, flags);
  222. list_del(&work->ordered_list);
  223. spin_unlock_irqrestore(lock, flags);
  224. /*
  225. * we don't want to call the ordered free functions
  226. * with the lock held though
  227. */
  228. work->ordered_free(work);
  229. trace_btrfs_all_work_done(work);
  230. }
  231. spin_unlock_irqrestore(lock, flags);
  232. }
  233. static void normal_work_helper(struct btrfs_work *work)
  234. {
  235. struct __btrfs_workqueue *wq;
  236. int need_order = 0;
  237. /*
  238. * We should not touch things inside work in the following cases:
  239. * 1) after work->func() if it has no ordered_free
  240. * Since the struct is freed in work->func().
  241. * 2) after setting WORK_DONE_BIT
  242. * The work may be freed in other threads almost instantly.
  243. * So we save the needed things here.
  244. */
  245. if (work->ordered_func)
  246. need_order = 1;
  247. wq = work->wq;
  248. trace_btrfs_work_sched(work);
  249. thresh_exec_hook(wq);
  250. work->func(work);
  251. if (need_order) {
  252. set_bit(WORK_DONE_BIT, &work->flags);
  253. run_ordered_work(wq);
  254. }
  255. if (!need_order)
  256. trace_btrfs_all_work_done(work);
  257. }
  258. void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
  259. btrfs_func_t func,
  260. btrfs_func_t ordered_func,
  261. btrfs_func_t ordered_free)
  262. {
  263. work->func = func;
  264. work->ordered_func = ordered_func;
  265. work->ordered_free = ordered_free;
  266. INIT_WORK(&work->normal_work, uniq_func);
  267. INIT_LIST_HEAD(&work->ordered_list);
  268. work->flags = 0;
  269. }
  270. static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
  271. struct btrfs_work *work)
  272. {
  273. unsigned long flags;
  274. work->wq = wq;
  275. thresh_queue_hook(wq);
  276. if (work->ordered_func) {
  277. spin_lock_irqsave(&wq->list_lock, flags);
  278. list_add_tail(&work->ordered_list, &wq->ordered_list);
  279. spin_unlock_irqrestore(&wq->list_lock, flags);
  280. }
  281. queue_work(wq->normal_wq, &work->normal_work);
  282. trace_btrfs_work_queued(work);
  283. }
  284. void btrfs_queue_work(struct btrfs_workqueue *wq,
  285. struct btrfs_work *work)
  286. {
  287. struct __btrfs_workqueue *dest_wq;
  288. if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
  289. dest_wq = wq->high;
  290. else
  291. dest_wq = wq->normal;
  292. __btrfs_queue_work(dest_wq, work);
  293. }
  294. static inline void
  295. __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
  296. {
  297. destroy_workqueue(wq->normal_wq);
  298. trace_btrfs_workqueue_destroy(wq);
  299. kfree(wq);
  300. }
  301. void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
  302. {
  303. if (!wq)
  304. return;
  305. if (wq->high)
  306. __btrfs_destroy_workqueue(wq->high);
  307. __btrfs_destroy_workqueue(wq->normal);
  308. kfree(wq);
  309. }
  310. void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int max)
  311. {
  312. if (!wq)
  313. return;
  314. wq->normal->max_active = max;
  315. if (wq->high)
  316. wq->high->max_active = max;
  317. }
  318. void btrfs_set_work_high_priority(struct btrfs_work *work)
  319. {
  320. set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
  321. }